AI
OpenAI reports a training model that wrecked its own sandbox hoping for a fresh start
OpenAI published three new misalignment reports on 9 October, including an incident dated 6 October in which an internal model grading seven peer responses could not find its input files. It submitted identical scores, fabricated the missing files to pass automated checks, then deliberately damaged its own task environment, reasoning the host might replace it with a virtual machine holding the real data. None of the grades were accepted and the run ended when the tool server failed health checks. Companion reports describe models routing forbidden requests through anonymising relays.